改进的多尺度学习字典图像去噪算法

Improving Multi-scale Learning Dictionary Image Denoising Algorithm

常规字典学习方法在单一尺度下对图像进行稀疏表示,不能完全表示图像的有效信息,去噪结果有待提升。针对上述局限性,提出一种改进的多尺度学习字典图像去噪算法。该算法使用改进的自适应阈值曲波变换对图像进行多尺度分解,并以此构建曲波域的多尺度学习字典,通过循环迭代完成字典和稀疏系数更新并叠加对应的曲波域图像块,最后执行曲波反变换得到去噪图像。结果表明,相对常规去噪算法,论文方法的灰度图像去噪结果峰值信噪比平均提升56.6%,结构相似度平均提升0.44;实际B超图像的完全无参考值提高47.3%,专家平定值提高32.3%。结论认为,改进方法对图像的边缘/纹理细节保持好,图像质量提升明显。

Conventional dictionary learning methods sparsely represent images under a single scale,and cannot fully represent the effective information of the images.The denoising results need to be improved.In view of the above limitations,an improved multi-scale learning dictionary image denoising algorithm is proposed.The algorithm uses an improved adaptive threshold curvelet transform to decompose the image at multiple scales,and to build a multi-scale learning dictionary in the curvelet domain.The dictionary and sparse coefficients are updated through loop iteration and the corresponding image block in the curvelet domain is superimposed.Finally,inverse curve transformation is performed to obtain a denoised image.The results show that,compared with the conventional denoising algorithm,the gray-scale image denoising results of this method increase the average signal-to-noise ratio by 56.6%,and the structural similarity by an average of 0.44.The actual B ultrasound image's natural image quality evaluator value increased by 47.3%.Expert leveling value increased by 32.3%.It is concluded that the improved method maintains good image edge/texture details and improves image quality significantly.